Enhancing U-Net-Based Super-Resolution Models for Real-Time Rainfall–Runoff Flood Simulation

Abstract

Real-time flood forecasting requires efficient high-resolution simulations. We combine the TRITON hydrodynamic solver with U-Net-based super-resolution models trained on paired coarse–fine simulations. A sensitivity analysis of different loss functions is conducted to optimize reconstruction accuracy and flood extent prediction, achieving 4× spatial enhancement with centimeter-level errors under realistic rainfall events.